Welcome to Episode 7 of Women WithAI, where we are thrilled to have Dr. Catherine Breslin, a leading AI consultant and machine learning scientist, join us to unpack the complexities of AI technologies and the gender disparities in the tech world.

In this conversation, we'll examine how biases enter our AI tools, examining everything from voice recognition to language models.

Dr. Breslin will share insights from her remarkable career, shedding light on her hands-on experiences and leadership roles. She'll discuss everything from algorithmic biases to the potential of multi-modal AI models. She'll also highlight the critical need for diversity in tech, emphasising strategies to encourage more women to enter and excel in this field.

By the end of our discussion, we'll better understand the significant role human oversight plays in shaping ethical AI and the importance of challenging norms to make technology inclusive and equitable for everyone. Stay with us.

Hello and welcome to Women WithAI, a podcast focusing on the challenges

and successes of women in this rapidly evolving sector. Today, I'm thrilled

to welcome Doctor Catherine Breslin, who is an AI consultant and machine learning

scientist with over two decades of experience as an AI scientist building

voice and language AI models. Catherine is the founder and director of

Kingfisher Labs, where she works with business leaders to bring cutting-edge technologies to

market. Her previous roles include AI scientist and manager at the

University of Cambridge, plus organisations such as Toshiba Research,

Amazon, Alexa and Cobalt speech. She's also an AI

advisor and coach and has been named one of Nesta's twelve women shaping AI

as the expert in machine learning. She was on the 2021 list of

Computer Weekly's 50 most influential women in UK tech.

Catherine Breslin, welcome to Women WithAI.

Thanks Jo, thanks for having me. It's lovely to have you here.

So how I'm going to start off and ask you how you got into doing

what you're doing. And for those that don't know, when it comes to

AI and machine learning, what is it? What's the difference? What is machine

learning? Fantastic. I'll start with maybe

talking a little bit about what AI is, maybe, and what differences. You hear a

lot of terms being thrown about right now, so maybe we can

dive straight into that and start talking about those. So AI,

obviously a term that's been in the media loads lately, and I think you'd have

to be hiding under a rock not to have read something about the technology

lately. And AI is a term that's been around a long time. It's gone in

and out of fashion as the technology has evolved and then lived up to

its promise or not lived up to its promise. And we're going through a phase

right now where AI technology is really making leaps

and bounds in performance. And so people are really excited about the

potential. So AI is a term which doesn't really

have a really great crisp definition, and people

use it to mean a lot of different things. And so in general, maybe we

can think about it as technology that is

trying to sort of emulate some sort of human decision process,

human decision making. So we're trying to automate things

which people can do which require a little bit more intelligence than just

following a list of instructions or a list of rules. So if you've done

any computer programming in the past, you'll know that what you do is you sit

there and you very carefully write out a list of rules, instructions for

computers to follow, to do something and so you can get quite far with that

sort of technology. You can make computers do quite a lot of things

by sitting there and writing down the rules for how to do it. But

there are some things that you really just can't write down the rules for. So

something like understanding speech, everyone's speech is different,

and everyone says different things. And so just writing down the

rules about how to understand what someone is saying, you know, it's just

impossible. There's so much context and variation that goes

into it that it would be an impossible task to write it down. And so

that's where we start to talk about machine learning. Machine learning is

a subset of AI, a group of algorithms that really

learn what to do from looking at data, from looking at examples.

So in that example of speech recognition and understanding

speech, we show the machine lots of examples of people

speaking, and we also show them the words that that person said, so they can

learn the patterns, learn the correlations, and understand, and

maybe learn a bit more about how people speak so that they can then go

on to transcribe other speech. And this idea of learning from

data, machine learning, and it's really what's been

driving this past decade of progress in AI, I think. And then when

you hear about AI, a lot of that is down to

machine learning and improvements and changes in machine learning.

Okay, great, because that explains it well, because I

guess as a human, you see how someone says something, you can

see the look on their face or the way they say it. So

can machines learn that as well? What about sarcasm?

Of course. Yeah. When you're talking to somebody, you can see a lot more. You

can see, like you say, their face. You have a shared conversation

history. You have some context and some cultural knowledge in there as well.

And whether machines can learn all of that at the

moment, I think they can't right now. I think

the amount of data it would take to sort of understand all of that cultural

knowledge and all of the context and nuance that goes into

speech and language. We're not really at the point that computers can get all

of that just yet, but we have made some strides in the past few

years in being able to build these

systems from bigger and bigger sets of data. And those bigger and bigger sets of

data do have a lot more sort of context and nuance in them.

So we're making steps in that direction. We still got some way to go.

I guess it's like being a child, isn't it? As a young child, you're not

going to get all the nuances or understand if someone isn't really

meaning what they say. So I guess maybe we're at the, at the beginning,

although maybe, and children. Are very good at

understanding speech. Yes. I've got two nieces, and it's

amazing. You know, they're seven and ten at the moment, and it's amazing. Yeah, they

definitely can pick up now while they're getting better and better, what it

means. But so how did we. Yeah, how did we get here, do you think?

Because OpenAI hasn't been an overnight success, as you say. But is

it that sort of, that exponential, that curve? Do you think it will suddenly start

to move a lot quicker? I think what has

happened in the past, maybe 15 years or so

now is that we've seen a few things that have come together to

make this technology more able to build the

capability that we're seeing today. So one of the first things has

happened is that the cost of computation, the

amount of computation you can do, the processing capability for

computer chips has got much better in the past decade or so, which allows

us to do a lot more on those chips. We also have the

Internet. The Internet has provided a place for people

to write a lot of text data, which is readable by

computers. So 1020 years ago, we didn't have just the

sheer amount of writing and audio and video on the Internet, as we

do now. So the amount of data available to companies

needs to build their models from has got a lot larger in the past few

years. And, of course, there is lots in the press right now about sort

of copyright and consent of using data. But a lot

of the large companies are using quite a lot of data from the web to

train their models, and that gives us much, much larger data sets.

And that's been another thing that has fed into the models, being more

capable as they are learning from more and more data, they will

understand much more of the nuance and learn many more,

much more context than when you had training these models

on small data sets. And I think the other thing that's happened

as well is that we've had some improvements in the underlying algorithms that

we use. So if you're following the field, you might have heard of sort of

deep learning and transformers diffusion models. Some of these techniques

have come along more recently, and they're able to model some of the

language and speech and audio better than our previous generation of

models. So these three things have really come together. So amount of data, amount of

computation, the Internet holding it all together, and the improved

algorithms, and that's really driven what we've seen in the progress in the

past, probably 15 years or so. Because it

is all about the data, isn't it, as you say? And I suppose it's learning

from the data that's already there. So, I mean, people, how can we

go about making sure that it's the right data? Or do we need, is there,

is there bias in the data? And in AI voice technology, for

example, are there any biases that you've come across?

Exactly. And bias is a big topic right now as well, I think, because

we're starting to see that if you do train

models on some of the larger data sets that we

see machine learning models, because they're learning patterns in data,

they learn whatever is there. They're not making conscious

decisions about whether something is biased or not, and they should

use it like humans sometimes do. They are just learning. Everything

is equal in that data set. And as you can imagine, quite a lot

of the writing, quite a lot of the speech on the Internet does exhibit certain

kinds of biases, and those biases then do just

sort of transfer straight through into our machine learning models that we are building.

And companies are putting a lot of effort into mitigating some of these biases now.

But I think some of the ways that you see it play out are with,

especially when we're thinking about voice and language technology, we see

a lot of different accents in the world.

So everybody, even here in the UK, we have so many different

accents, but only some of those accents are

better recognized by computers than others. So we sort of see

this uneven distribution of performance across different accents

is one way we see this. We see

this technology being developed much more for languages

like English, Spanish, Mandarin, for which there is lots of data. And

of course, there are something like six and a half thousand languages in the world,

and very few of those have enough written data

to be able to build the same level of model from. So we

see an uneven distribution in the languages that we're

covering as well. So something like English,

very much more capable technology than some of these,

what we call sort of low resource languages. So we see

different aspects like this in voice technology,

moving on to language technology as well. And people talk about the

biases. If you've played with any of these language models, say chat,

GPT or Claude, or any of

these models, they are trained on data which exhibits the

views of the Internet, which is also very

western biased, and exhibits a lot of racial and gender biases

as well. And those can carry through into the models. When you, you

start to train on them. So I think there's different ways,

different in these different places, that some of that bias comes

through to the technology, different challenges. And that, I suppose,

leads me on to thinking, you know, I mean, I know other voices are available

and you can choose the voice of your AI, but most AI voices, well,

to me, anyway, you know, including Alexa and Siri, are female voices.

Why do you think that is? Yeah, this is another

way that I think we see some of society's bias play out

in technology. So when these systems were

built, probably, I don't know, 1015 years ago, Alexa series,

a lot of these voice assistants were built. It was much more

difficult to build a synthetic voice. It took a lot of effort

to record audio from one

person and convert that into a synthetic version of their

voice. So it was very time intensive, very expensive to build multiple

voices. So a lot of these

organizations, a lot of these projects started with the idea of offering a

diversity of voices to people, but realized that practically

it was very difficult to build them, and so they sort of settle on

one voice. And there is a lot of

evidence in the literature that people tend to prefer female

voices as well. And so you see this reinforcing cycle where

organizations will choose voices that people refer, which reflect the biases

in society and sort of embed and entrench those. And

so the cycle sort of continues. Now, I think we're in a

situation where companies, synthetic voices, it's

much easier to make them in a variety of voices, and companies are starting now

to offer a lot more variety in the voice that they do. But some of

this people still associate the voice assistance with

female voices. I mean, it could be. I mean, I've

done a bit of reading around the subject, and is it because female voices are

maybe less threatening because it's sometimes it's

easier to sort of have, I don't know, a female in the role

of assistant. I mean, I don't know, they're the biases that you don't want to

encourage, do you? Is it, was it

someone, a friend said to me the other day, is it due to Star Trek?

Is it because Star Trek, when they had the computer, it was a female voice?

And it's just, it sort of all started from there. Or then you look onto

films and tv and. But generally,

robots tend to look female. But is that because

they're being designed by males, or are they being designed by

females? Or is it because they're. They're just less scary than a, you know, a

terminator, like the male version of the robot?

I'm not sure. Like, how do we make them gender neutral? Or should we?

Yes, an interesting question. People have tried, I have seen sort of a

gender neutral voice that people have developed. But one of

the things that even, no matter how neutral, you try and make a voice, you

know, people. I still found myself making assumptions about the person

behind that voice. So every voice. There is

really no sort of neutral voice. Every voice has some sort of

cultural or, you know, associations with it that.

And therefore. So I think my view is that we want to offer variety

rather than, you know, try and build a neutral voice,

because there is no such thing as a neutral voice. Like, there's no such thing

as a neutral accent. Although a lot of people feel like, I don't have an

accent. Yeah, I was. I don't have an accent.

Yeah, but it depends where you are. But I found as

well, my Alexa, I hope

she's not listening. She might start speaking. But I have it slightly

speeded up. And I know that's something that when other people come around to my

house and they. They ask something and what's wrong with her? And I said, oh,

I just had it speeded up quickly. You know, if I need to know what

the weather's like in the morning, I haven't got time. I need to know straight

away. Tell it to me quickly. And I've tried

different, you know, having the different voices, but I quite liked it, you know, playing

with it when I had Siri to begin with, and I said, oh, I quite

like watching neighbours, maybe I'll have the australian voice. And I

found that I preferred the australian woman to the australian man because

it was easier to understand. But, yeah, I've gone back to the british version

now, but talking about bias and

women in industry, I know that you're keen to get more women and girls

interested in STEM, and you co founded the Cambridge branch of the British

Science association and Robogals. So can you tell our audience a little

bit about that, please, and how we can get more girls

interested in STEM? I mean, we do

have. There's a big gender problem in technology, and maybe

the same applies for racial bias and other sort

of minorities as well. But there are really a lot of statistics out there

that show us that women are not choosing to work in

technology, and if they are, they are not sort of rising

up the ladder and making it into senior positions and being some of those

leaders in the field. So maybe here in the UK, we know

that it's quite difficult to get an exact figure, but around about 20%

of the AI workforce is women. And we also know

that when it comes to

funding of startups, and a lot of startups are AI focused at the

moment. Funding of startups, all women teams, the last figures I

saw got less than 2% of the venture capital funding,

compared to all male teams who got something like 80% of the

funding, and mixed teams got the rest. And then we know there's a gender pay

gap. We know that women don't make it into leadership positions at the same rate

as men do. So all of these show us that there really is a problem

with women in technology and women in AI. And there are

two ways, I think, to think about this. So the first one

is that, you know, encouraging young women and girls into

the field in the first place, and the second is sort of

promoting and appreciating the women that are already there and providing career

paths and bringing those up into senior positions. I think both are

important, getting girls

interested in science and technology. And things are changing. I think, as

you start to see the impact of technology in society a bit more, I think

a few more girls use a lot of apps, use a

lot of social media, sort of start to understand the impact and have a little

bit more interest perhaps in the computer science behind that, but still

not choosing to go on and study sort of computer science and maths and technology

and engineering subjects at university. So I think

encouraging women there and showing them from a young

age, I think people start to make their decisions about which

subjects boys and girls are good at. So really going into primary schools

and early secondary school and showing them that this is a valid career for people

to take, and there are women already in this field to look up to.

But I do think that without the second factor,

without sort of encouraging the women that are already there, so companies

have, women do not tend to make it into leadership

positions in general, but also in technology at the same

rate as men. And that's because

of various factors. But I think sort of maternity and

motherhood is one important place where women of young kids

really need flexibility and employers do not offer it to them

in the way that they need. And so that forces a lot of women to

sort of take a step back or to drop out of the workforce for a

little while and really, you know, holds them back. So flexibility, I

think, is really important here. And also there's a lot of

then bias in what a leader looks like and whether women

can be promoted into those positions. And, you

know, when you get higher up and there are fewer and fewer women, it

becomes harder and harder to get promoted up. And so I think

paying attention and really noticing

the pay gaps that you have in your company and the way that you are

treating your women leaders and the way that you are bringing up their careers, I

think all of that is really important in equalising

the field. Yeah, definitely. So, yeah, it's not just getting people

interested in it, it's getting people coding, it's keeping them doing

it. Yeah, because I think one of the, like,

young girls and women, it's great to get them into the field. I think it's

a great career, but they are not the ones with the power to change it

in the long run. You need to, you need the leaders to be the ones

changing the field. And so getting more women into leadership positions

is really the only way. Exactly. Getting them in there so they can make the

change. And it's flexible working. I mean, that's the

hope, isn't it? I mean, with AI, lots of it, you know, it's talking about

these tools and how can they make everything easier. So we needed to, to really

start doing that and making sure if women are leading it, then

using it for how we can get people to stay in industry.

Because do you, do you build any products? Because I've seen you

talking for about large language models and advising people how to use it in business

and risks and opportunities. Have you been involved in actually kind of

building those products?

So I have. Earlier in my

career, I was a sort of hands on computer programmer, sitting down,

writing the code for some of these things, building the models that went into

some of these products, and some of the research that we have

been making up the field.

That was my early to mid career. And

about a few years ago, I moved into more of a management position,

so I started managing people. I moved a little away from hands on coding

more onto some of the strategic and leadership

thinking around AI, what we should be building. Now I

work with companies who are building technology and

I'm hoping to form the bridge between what's going

on in the research world and what the technology, how the technology is developing and

keeping abreast of all those and figuring out how that can be used and how

that can be incorporated by companies in their work. Okay,

great. And I've seen you do a lot of sort of best practice and

that kind of thing. What does best practice in AI look like? What does that

mean? Oh, I think we would need a whole other

podcast for best practice in AI. I think we're

still really trying to figure this out a little bit because AI is such a

new field and we're kind of making up what the best

practices at the moment. A lot of people think that building

an AI product is about building the thing and getting it out into the world,

but really that's only a part of the job. We have

to then look after that product and improve it and make it better over

time. And that ends up being a much bigger part

of the job, I think, than people realize when they start out. So putting

in place ways

to monitor what your product's doing, that's really

important to know how well it's doing in the real world and

not to put something out there which kind of works in the lab, but doesn't

really work in the real world. We talked about bias already,

so looking out for biases, trying to mitigate them at

the early stages, we didn't touch on another part of

bias, which I think is the sorts of decisions about what you build and the

products that you are building, who they're aimed at, and whether

they actually fulfill the needs that people have. So

choosing what you're going to build, I think very thoughtfully, is also a big part

of this. And then I think a really

big part of best practice is just testing, properly testing and

evaluating. It's really such a big part

of building an AI product is making sure that it works like you think it

does, being able to justify it and being able to know

when it works and also when it doesn't work so that you can mitigate some

of that and you can have a person in the loop to

work with the system when it's not working or to understand

its strengths and weaknesses as well, because. I guess that comes on to sort of

like regulations and stuff. Like is anyone regulating it or is it very much up

to each organization or whoever's deciding to build it? Is there

any kind of regulation out there? Is there regulation?

Yes, yes, regulation. Another big. Like, lots of the topics around

AI are really big right now. Regulation is another one.

Maybe a month ago, the EU signed into being the EU

AI act, which I think is the first real

regulation of AI technology that splits

AI technology into different risk categories. So there's a sort of

unacceptable risk, a high risk, a medium risk, low risk categories,

and there are different obligations at different levels of that technology.

So unacceptable risk technology is banned,

whereas high risk technology has more obligations

associated with transparency and reporting with it. And then low

risk technology is, has a much lower

bar in the regulation. So that's one example of regulation which has

come into being. Other companies are thinking, other countries are

thinking about regulations and of course there are

associated regulations that are already in place. So things like

the EU's GDPR and

some of the health regulation that exists

in Europe and the US for health devices and health data,

some of the financial regulation, if you're working in the financial.

So there's sector specific regulation as well, which does touch

on AI technology. And do you think it's always people,

isn't it, that are checking that and doing the regulating. So

AI is not quite taken over the world and deemed as useless yet still.

Well, it needs us because we're building it. It's our product, isn't it? So

got to keep an eye on it. And how do you use AI, like in

your personal life? How do I use AI in my

personal life? That's a good question.

I tend to use it at work. You're like

a chef that cooks and goes home and doesn't want to, doesn't want to do

any of the cooking.

I think we have AI weaved into our everyday lives in

ways that are not always noticeable. So

things like I'm a photographer as well,

so I have a large library of photos. You can search photos now

for, you know, objects or people

and we all go on our phone and hopefully, you know, you can

see the phone has categorized all the photos that have a picture of me

or a picture of my family in. So

those things are really helpful. You can, you can search within your image library

to find a particular picture. If I know I took a picture of, I don't

know, a rainbow five years ago and I want to find that picture, I

can search for rainbow and it will bring up the pictures in my library. So

one way that AI is used that, you know, we sort of maybe start to

take for granted now, but where we have behind the

scenes companies building models to understand what is in images, to help

us look through them. And, you know, this is really important. If we

have smartphones that take lots of pictures and we start to take many, many more

pictures than we used to have and they're just sitting there on our phone,

difficult to work through. So I think that's one

way that people see AI in their daily lives.

A lot of people included use chat, GPT for various

different things. It's a great helper, brainstormer and rephrase

things sometimes, things like that. I know

there's a lot of people in the AI world now looking

at some of these new code tools, so

understand tools that will help you write computer code

proving to be quite useful. I think if you know what you're doing and if

you know how to use them, you can be much quicker

at writing and brainstorming and getting your computer code written. And that can be

really helpful for just day to day work

if you're doing that. So I think there's lots of different ways. It's not like

a sort of big, it's just every thing that you're using,

but just little things in your life where AI crops up. Yeah.

Exciting. And I read some, I heard something the other

day that, about the sort of the picture learning thing. I don't know if you

can, if you know if this is true or not, but you know, when you

have the, the captcha thing, so you go on a website and you've put in

your data or it's asking for something and it checks, you know, are you a

robot? And so you have to choose, you know, how many of the six or

eight squares or however many squares, nine squares have got a picture of a traffic

signal or a bus or a cat or a dog or something like that? Is

that helping the machine learning? I mean, is that, or is that just,

I don't know, is that just something that was

invented and we all do? Or somehow does that data get fed

into AI? Yeah. So

at the beginning we talked about how to teach a machine

to understand speech. And I said that you had audio with the

transcription, so you know what was said in the

audio. To build a speech system, if you're building a system which is going to

know something about images, you need the same sort of idea for images. So you

need images and you need associated text or

labels or something to tell you what's in the

image. And so usually what we have

is people who annotate those images, annotate the audio,

maybe annotate the text, data, whatever it is that you're looking at with the

correct answer, what's actually in that image or piece

of audio or what's in that text file.

And so getting those labels

for different images, people have very creative ways to do it. And so

capture the, you know, click which images have a bridge in or which

images have a traffic light in is one way to get some

of those human verified labels for images.

Cool. And how do you see AI?

Sort of. What are you excited about over the next, you know, like

2345 years, what do you see as the sort of big advancements or what

would you like to see happening?

So what would I like to see? I think we're at a really interesting point

right now because we had sort of chat

GPT launched two years ago. This was a big turning point

in public awareness. They think of AI technology and what it could

and couldn't do. And now we start to see similar

models that are able to deal with not just text, but images. I

mean, GPT nowadays will deal with images as well if you pay for the

subscription, audio, video. We're starting to see all these things come together,

which I think is really interesting because that's going to allow

many different capabilities. I think that's one of the useful

things people find about chat GPT, is its ability to use image as well as

text now. So this sort of multimodal models, I

think are really interesting. And we're in a world

where open source technology is also

building many of these things. I think open source is really interesting because

then we can build these models and

lots of people can try using this. It's very expensive to build the model in

the first place. To build GPT, chat GPT or to build.

There's a model that Facebook or meta launched called Llama.

There's other models that other people have launched. These are very expensive to build,

but once they're built, people can take them and run with them and try

them in their own domains and own fields. So I think open source is really

helping with this. And we see, you know, in

scientific research, for example, researchers come up with really creative

ways to try and use these, what we call foundation models or

frontier models to build

on for their own specific domains. So I'm really interested to

see what people do with them and what

they find they're capable of. Also trying to just figure out

what these models are and aren't capable of. We've had a couple of

years of experimentation and people have found some really interesting things that they can

do, and I think that will continue as well. So I think we will see

a big explosion of this technology used across a wide

variety of domains, where you've got domain experts who know about their

field, working with these models that have been built by open

source communities or big organizations with the money to do

so. That's really exciting. Fantastic. So there's so many

opportunities out there, aren't there? So yeah, just need to embrace

those. So where can our audience find out more about everything you've

done? And are you available to go into primary schools? And if people want you

to do that, how can they get in touch? I

have been into primary schools and I'm very happy to do so. I

think LinkedIn is the best place to find me. If you're interested to know, LinkedIn

and I have a website, we'll put links. To that in the show notes. Because

you've got. You do a newsletter as well, don't you, on substack? And another

writing is because. Yeah, you. I like one of your

other accolades that you had, was one of the hundred coolest people in the

UK tech world. So it's been

fantastic to speak to you. Katherine, thank you so much for coming on. You've got.

There's so much more we can talk about. We'll have to get you back on

and talk about best practice and regulation.

Fantastic. Really great to speak with you. Thanks

for coming.